01
Scope behaviours, not scripts
The rules the agent must never break come first: what it must always say, what it must never claim, when it must get out of the way.
Telephony · Speech · Safety
Voice AI Agents Safe Enough for a Real Phone Line
Anseya, our AI phone receptionist, answers real inbound calls in production. Building it taught us what a voice agent actually requires: speech recognition that survives accents and language switching mid-sentence, disclosure by rule rather than by mood, and emergency handling that is deterministic code, never a model's judgement call. We build voice agents with those lessons already paid for.
What you get
Deterministic safety
Emergencies are detected by a phrase lexicon over the live transcript, before any model is consulted. A caller who smells gas hears a scripted response directing them to the gas emergency line, immediately and unconditionally.
Disclosure by rule
The agent introduces itself as an AI assistant and states that the call is recorded, on every call. Honesty is not a configuration option.
Multilingual that is real
Callers switch languages mid-call. We engineer for it with code-switching speech recognition, and a language ships only when its emergency phrasing is covered too.
The full pipeline
Telephony, speech-to-text, the reasoning model, text-to-speech and your business tools, wired together and measured end to end for latency.
Proof you can open
Not case studies you have to take on trust. These are live, and you can use them before you ever talk to us.
How we deliver it
02
Wire safety first
The emergency path is built and tested before the sales path, because the call that matters most is the one where someone is in danger.
03
Choose the stack by measurement
Speech engines benchmarked on your real call audio at telephone quality. We changed our own transcription stack when the measurements said so.
04
Rehearse on real calls
Live test calls before launch. Our own first six real calls surfaced eight faults no offline test had found; yours get that treatment too.
Tools we use
VapiDeepgramAzure SpeechClaudeFastAPISupabase
Frequently asked
Does the caller know they are talking to an AI?
Yes, always. Anseya opens by introducing itself as an AI assistant and disclosing that the call is recorded, and agents we build for you do the same. Beyond honesty, disclosure defuses the frustration of a caller who discovers it mid-call.
What happens if a caller reports an emergency?
A deterministic detector watches the transcript for emergency phrasing: gas, fire, electrical, flood, medical. It fires before any AI generation, and the caller hears a scripted safety response, such as being told to hang up and call the gas emergency line. Safety information is never delayed for lead capture and never gated behind a plan.
Can it speak our customers' languages?
Voice agents can work across languages, and Anseya answers in the caller's own language. We hold a hard rule you should demand from any vendor: a language is enabled only when the emergency safety phrasing is covered in that language too, not just the sales flow.
What does a voice agent cost to run?
The per-minute cost stacks telephony, speech-to-text, the language model and speech synthesis. We engineer each layer for its job, measure the stack end to end, and show you the per-call economics before you commit.
Related reading